Skip to content

voicegrid-operations.boundary-model.md

Latest commit

 

History

History
74 lines (53 loc) · 3.32 KB

voicegrid-operations.boundary-model.md

File metadata and controls

74 lines (53 loc) · 3.32 KB
Raw

Voice And Transcript Boundary Model

Purpose

This profile-testbed example describes the agentic boundaries for VoiceGrid Operations.

It is not a repo-local production claim. The repository remains independent: it is not an AAIF project, not a Linux Foundation project, and not endorsed by either organization.

High-Risk Action Boundary

Voice generation, transcription, speaker attribution, consent tracking, transcript export, and provider selection for speech workflows.

Actors

Actor Responsibility
Operator Approves voice jobs, transcript handling, and export decisions.
Speaker or subject Owns consent-sensitive voice and transcript material.
Media agent Proposes transcription, synthesis, summarization, and routing actions.
Speech provider adapter Executes STT/TTS calls through configured providers.
Maintainer Reviews evidence, fallback providers, and privacy-sensitive changes.

Identity Boundary

  • Voice and transcript actions must be associated with a human operator.
  • Speaker consent must be represented separately from operator identity.
  • Provider credentials belong to the deployment environment, not the agent.
  • Generated media must carry enough provenance to distinguish source audio, transcript, and synthesized output.

Tool And Action Boundary

Action Boundary
transcribe_audio Allowed only for uploaded or linked media with consent or policy basis.
synthesize_voice Requires approved text, voice profile, and output-use context.
summarize_transcript Must preserve source transcript reference and confidence limits.
export_media Requires human approval and redaction review for sensitive content.
switch_provider Requires provider capability, cost, privacy, and fallback review.

Human Approval Gates

  • Before synthesizing a voice that resembles a real person.
  • Before exporting transcript or generated audio outside the workspace.
  • Before using a paid or hosted speech provider for sensitive material.
  • Before retaining media beyond the documented retention window.

Audit Event Model

Event Minimum Evidence
media.uploaded operator id, source, consent basis, checksum, timestamp
transcript.created provider, model, language, duration, confidence pointer
voice.generated provider, voice profile, text hash, approval id
media.export_requested target, redaction status, approver, timestamp
provider.changed previous provider, new provider, reason, reviewer

Local Proof Path

  • npm install to prove dependency resolution.
  • npm run lint to prove static project health.
  • npm run test for media workflow and policy checks when available.
  • npm run build to prove the application compiles.
  • Use sample media only; do not include real customer or speaker data in public proof.

OSS And Self-Hosted Fallbacks

  • Use local Whisper, faster-whisper, or Vosk for STT trials where feasible.
  • Use Piper, Coqui-style engines, or other self-hosted TTS paths for non-sensitive experiments.
  • Keep paid providers behind adapters with explicit environment variables.
  • Store audit events in PostgreSQL or append-only logs before relying on hosted-only telemetry.

Contributor Follow-Up

Add a repo-local docs/AGENTIC_BOUNDARY_MODEL.md after review and link it from agentic-readiness.json.